The handling of the no-preemption priority level imposes the restriction
that we need to maintain the implied ordering even though preemption is
disabled. Otherwise we may end up with an AB-BA deadlock across multiple
engine due to a real preemption event reordering the no-preemption
WAITs. To resolve this issue we currently promote all requests to WAIT
on unsubmission, however this interferes with the timeslicing
requirement that we do not apply any implicit promotion that will defeat
the round-robin timeslice list. (If we automatically promote the active
request it will go back to the head of the queue and not the tail!)
So we need implicit promotion to prevent reordering around semaphores
where we are not allowed to preempt, and we must avoid implicit
promotion on unsubmission. So instead of at unsubmit, if we apply that
implicit promotion on adding the dependency, we avoid the semaphore
deadlock and we also reduce the gains made by the promotion for user
space waiting. Furthermore, by keeping the earlier dependencies at a
higher level, we reduce the search space for timeslicing without
altering runtime scheduling too badly (no dependencies at all will be
assigned a higher priority for rrul).
v2: Limit the bump to external edges (as originally intended) i.e.
between contexts and out to the user.
Testcase: igt/gem_concurrent_blit
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190515130052.4475-3-chris@chris-wilson.co.uk
Currently there is an underlying assumption that i915_request_unsubmit()
is synchronous wrt the GPU -- that is the request is no longer in flight
as we remove it. In the near future that may change, and this may upset
our signaling as we can process an interrupt for that request while it
is no longer in flight.
CPU0 CPU1
intel_engine_breadcrumbs_irq
(queue request completion)
i915_request_cancel_signaling
... ...
i915_request_enable_signaling
dma_fence_signal
Hence in the time it took us to drop the lock to signal the request, a
preemption event may have occurred and re-queued the request. In the
process, that request would have seen I915_FENCE_FLAG_SIGNAL clear and
so reused the rq->signal_link that was in use on CPU0, leading to bad
pointer chasing in intel_engine_breadcrumbs_irq.
A related issue was that if someone started listening for a signal on a
completed but no longer in-flight request, we missed the opportunity to
immediately signal that request.
Furthermore, as intel_contexts may be immediately released during
request retirement, in order to be entirely sure that
intel_engine_breadcrumbs_irq may no longer dereference the intel_context
(ce->signals and ce->signal_link), we must wait for irq spinlock.
In order to prevent the race, we use a bit in the fence.flags to signal
the transfer onto the signal list inside intel_engine_breadcrumbs_irq.
For simplicity, we use the DMA_FENCE_FLAG_SIGNALED_BIT as it then
quickly signals to any outside observer that the fence is indeed signaled.
v2: Sketch out potential dma-fence API for manual signaling
v3: And the test_and_set_bit()
Fixes: 52c0fdb25c ("drm/i915: Replace global breadcrumbs with per-context interrupt tracking")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190508112452.18942-1-chris@chris-wilson.co.uk
Asking the GPU to busywait on a memory address, perhaps not unexpectedly
in hindsight for a shared system, leads to bus contention that affects
CPU programs trying to concurrently access memory. This can manifest as
a drop in transcode throughput on highly over-saturated workloads.
The only clue offered by perf, is that the bus-cycles (perf stat -e
bus-cycles) jumped by 50% when enabling semaphores. This corresponds
with extra CPU active cycles being attributed to intel_idle's mwait.
This patch introduces a heuristic to try and detect when more than one
client is submitting to the GPU pushing it into an oversaturated state.
As we already keep track of when the semaphores are signaled, we can
inspect their state on submitting the busywait batch and if we planned
to use a semaphore but were too late, conclude that the GPU is
overloaded and not try to use semaphores in future requests. In
practice, this means we optimistically try to use semaphores for the
first frame of a transcode job split over multiple engines, and fail if
there are multiple clients active and continue not to use semaphores for
the subsequent frames in the sequence. Periodically, we try to
optimistically switch semaphores back on whenever the client waits to
catch up with the transcode results.
With 1 client, on Broxton J3455, with the relative fps normalized by %cpu:
x no semaphores
+ drm-tip
* patched
+------------------------------------------------------------------------+
| * |
| *+ |
| **+ |
| **+ x |
| x * +**+ x |
| x x * * +***x xx |
| x x * * *+***x *x |
| x x* + * * *****x *x x |
| + x xx+x* + *** * ********* x * |
| + x xx+x* * *** +** ********* xx * |
| * + ++++* + x*x****+*+* ***+*************+x* * |
|*+ +** *+ + +* + *++****** *xxx**********x***+*****************+*++ *|
| |__________A_____M_____| |
| |_______________A____M_________| |
| |____________A___M________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 2.60475 3.50941 3.31123 3.2143953 0.21117399
+ 120 2.3826 3.57077 3.25101 3.1414161 0.28146407
Difference at 95.0% confidence
-0.0729792 +/- 0.0629585
-2.27039% +/- 1.95864%
(Student's t, pooled s = 0.248814)
* 120 2.35536 3.66713 3.2849 3.2059917 0.24618565
No difference proven at 95.0% confidence
With 10 clients over-saturating the pipeline:
x no semaphores
+ drm-tip
* patched
+------------------------------------------------------------------------+
| ++ ** |
| ++ ** |
| ++ ** |
| ++ ** |
| ++ xx *** |
| ++ xx *** |
| ++ xxx*** |
| ++ xxx*** |
| +++ xxx*** |
| +++ xx**** |
| +++ xx**** |
| +++ xx**** |
| +++ xx**** |
| ++++ xx**** |
| +++++ xx**** |
| +++++ x x****** |
| ++++++ xxx******* |
| ++++++ xxx******* |
| ++++++ xxx******* |
| ++++++ xx******** |
| ++++++ xxxx******** |
| ++++++ xxxx******** |
| ++++++++ xxxxx********* |
|+ + + + ++++++++ xxx*xx**********x* *|
| |__A__| |
| |__AM__| |
| |__A_| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 2.47855 2.8972 2.72376 2.7193402 0.074604933
+ 120 1.17367 1.77459 1.71977 1.6966782 0.085850697
Difference at 95.0% confidence
-1.02266 +/- 0.0203502
-37.607% +/- 0.748352%
(Student's t, pooled s = 0.0804246)
* 120 2.57868 3.00821 2.80142 2.7923878 0.058646477
Difference at 95.0% confidence
0.0730476 +/- 0.0169791
2.68622% +/- 0.624383%
(Student's t, pooled s = 0.0671018)
Indicating that we've recovered the regression from enabling semaphores
on this saturated setup, with a hint towards an overall improvement.
Very similar, but of smaller magnitude, results are observed on both
Skylake(gt2) and Kabylake(gt4). This may be due to the reduced impact of
bus-cycles, where we see a 50% hit on Broxton, it is only 10% on the big
core, in this particular test.
One observation to make here is that for a greedy client trying to
maximise its own throughput, using semaphores is the right choice. It is
only the holistic system-wide view that semaphores of one client
impacts another and reduces the overall throughput where we would choose
to disable semaphores.
The most noticeable negactive impact this has is on the no-op
microbenchmarks, which are also very notable for having no cpu bus load.
In particular, this increases the runtime and energy consumption of
gem_exec_whisper.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Dmitry Ermilov <dmitry.ermilov@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190504070707.30902-1-chris@chris-wilson.co.uk
It used to be handy that we only had a couple of headers, but over time
intel_drv.h has become unwieldy. Extract declarations to a separate
header file corresponding to the implementation module, clarifying the
modularity of the driver.
Ensure the new header is self-contained, and do so with minimal further
includes, using forward declarations as needed. Include the new header
only where needed, and sort the modified include directives while at it
and as needed.
No functional changes.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/2e4fb1e67ed38870df3040bb0a1b1a58fd90cc86.1556540890.git.jani.nikula@intel.com
It used to be handy that we only had a couple of headers, but over time
intel_drv.h has become unwieldy. Extract declarations to a separate
header file corresponding to the implementation module, clarifying the
modularity of the driver.
Ensure the new header is self-contained, and do so with minimal further
includes, using forward declarations as needed. Include the new header
only where needed, and sort the modified include directives while at it
and as needed.
No functional changes.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/64e46278dc8dccc9c548ef453cb2ceece5367bb2.1556540890.git.jani.nikula@intel.com
We switched to a tree of per-engine HW context to accommodate the
introduction of virtual engines. However, we plan to also support
multiple instances of the same engine within the GEM context, defeating
our use of the engine as a key to looking up the HW context. Just
allocate a logical per-engine instance and always use an index into the
ctx->engines[]. Later on, this ctx->engines[] may be replaced by a user
specified map.
v2: Add for_each_gem_engine() helper to iterator within the engines lock
v3: intel_context_create_request() helper
v4: s/unsigned long/unsigned int/ 4 billion engines is quite enough.
v5: Push iterator locking to caller
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190426163336.15906-7-chris@chris-wilson.co.uk
Broadwater and the rest of gen4 do support being able to saving and
reloading context specific registers between contexts, providing isolation
of the basic GPU state (as programmable by userspace). This allows
userspace to assume that the GPU retains their state from one batch to the
next, minimising the amount of state it needs to reload and manually save
across batches.
v2: CONSTANT_BUFFER woes
Running through piglit turned up an interesting issue, a GPU hang inside
the context load. The context image includes the CONSTANT_BUFFER command
that loads an address into a on-gpu buffer, and the context load was
executing that immediately. However, since it was reading from the GTT
there is no guarantee that the GTT retains the same configuration as
when the context was saved, resulting in stray reads and a GPU hang.
Having tried issuing a CONSTANT_BUFFER (to disable the command) from the
ring before saving the context to no avail, we resort to patching out
the instruction inside the context image before loading.
This does impose that gen4 always reissues CONSTANT_BUFFER commands on
each batch, but due to the use of a shared GTT that was and will remain
a requirement.
v3: ECOSKPD to the rescue
Ville found the magic bit in the ECOSKPD to disable saving and restoring
the CONSTANT_BUFFER from the context image, thereby completely avoiding
the GPU hangs from chasing invalid pointers. This appears to be the
default behaviour for gen5, and so we just need to tweak gen4 to match.
v4: Fix spelling of ECOSKPD and discover it already exists
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20190419172720.5462-1-chris@chris-wilson.co.uk
Ironlake does support being able to saving and reloading context specific
registers between contexts, providing isolation of the basic GPU state
(as programmable by userspace). This allows userspace to assume that the
GPU retains their state from one batch to the next, minimising the
amount of state it needs to reload, or manually save and restore.
v2: Fix off-by-one in reading CXT_SIZE, and add a comment that the
CXT_SIZE and context-layout do not match in bspec, but the difference is
irrelevant as we overallocate the full page anyway (Ville).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190419111749.3910-2-chris@chris-wilson.co.uk
Despite what I think the prm recommends, commit f2253bd985
("drm/i915/ringbuffer: EMIT_INVALIDATE after switch context") turned out
to be a huge mistake when enabling Ironlake contexts as the GPU would
hang on either a MI_FLUSH or PIPE_CONTROL immediately following the
MI_SET_CONTEXT of an active mesa context (more vanilla contexts, e.g.
simple rendercopies with igt, do not suffer).
Ville found the following clue,
"[DevCTG+]: For the invalidate operation of the pipe control, the
following pointers are affected. The
invalidate operation affects the restore of these packets. If the pipe
control invalidate operation is completed
before the context save, the indirect pointers will not be restored from
memory.
1. Pipeline State Pointer
2. Media State Pointer
3. Constant Buffer Packet"
which suggests by us emitting the INVALIDATE prior to the MI_SET_CONTEXT,
we prevent the context-restore from chasing the dangling pointers within
the image, and explains why this likely prevents the GPU hang.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190419111749.3910-1-chris@chris-wilson.co.uk
In the current scheme, on submitting a request we take a single global
GEM wakeref, which trickles down to wake up all GT power domains. This
is undesirable as we would like to be able to localise our power
management to the available power domains and to remove the global GEM
operations from the heart of the driver. (The intent there is to push
global GEM decisions to the boundary as used by the GEM user interface.)
Now during request construction, each request is responsible via its
logical context to acquire a wakeref on each power domain it intends to
utilize. Currently, each request takes a wakeref on the engine(s) and
the engines themselves take a chipset wakeref. This gives us a
transition on each engine which we can extend if we want to insert more
powermangement control (such as soft rc6). The global GEM operations
that currently require a struct_mutex are reduced to listening to pm
events from the chipset GT wakeref. As we reduce the struct_mutex
requirement, these listeners should evaporate.
Perhaps the biggest immediate change is that this removes the
struct_mutex requirement around GT power management, allowing us greater
flexibility in request construction. Another important knock-on effect,
is that by tracking engine usage, we can insert a switch back to the
kernel context on that engine immediately, avoiding any extra delay or
inserting global synchronisation barriers. This makes tracking when an
engine and its associated contexts are idle much easier -- important for
when we forgo our assumed execution ordering and need idle barriers to
unpin used contexts. In the process, it means we remove a large chunk of
code whose only purpose was to switch back to the kernel context.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190424200717.1686-5-chris@chris-wilson.co.uk
Start acquiring the logical intel_context and using that as our primary
means for request allocation. This is the initial step to allow us to
avoid requiring struct_mutex for request allocation along the
perma-pinned kernel context, but it also provides a foundation for
breaking up the complex request allocation to handle different scenarios
inside execbuf.
For the purpose of emitting a request from inside retirement (see the
next patch for engine power management), we also need to lift control
over the timeline mutex to the caller.
v2: Note that the request carries the active reference upon construction.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190424200717.1686-4-chris@chris-wilson.co.uk
We wish to start segregating the power management into different control
domains, both with respect to the hardware and the user interface. The
first step is that at the lowest level flow of requests, we want to
process a context event (and not a global GEM operation). In this patch,
we introduce the context callbacks that in future patches will be
redirected to per-engine interfaces leading to global operations as
required.
The intent is that this will be guarded by the timeline->mutex, except
that retiring has not quite finished transitioning over from being
guarded by struct_mutex. So at the moment it is protected by
struct_mutex with a reminded to switch.
v2: Rename default handlers to intel_context_enter_engine.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190424200717.1686-3-chris@chris-wilson.co.uk
